Abstract
Mass locust invasions pose a significant threat to agricultural productivity in Kazakhstan. Locust attacks pose a serious threat to Kazakhstan's agricultural productivity. In order to determine the fundamental predictors of the number of locusts in the Central, Eastern and northern parts of Kazakhstan for each agro-climatic zone, on the basis of machine learning algorithms, perennial historical data on the dynamics of the number of studied pests in the period 2003-2024, depending on the influence of weather, climatic and edaphic (soil) factors, were processed and analyzed. Agroclimatic zones with different characteristics were considered. Such a wide coverage of the regions made it possible to take into account thediversity of conditions and identify the main factors (main predictors) that affect the number of phytophages studied. In this study, using field data from 2003 to 2024 and regional climate variables from ERA5 obtained via Google Earth Engine, we applied machine learning models to forecast the population dynamics of migratory locusts.We employed CatBoost, SARIMAX, and LSTM algorithms, comparing their performance on the input datasets. The LSTM model achieved the highest prediction accuracy (MAE = 0.36, RMSE = 0.59), effectively capturing temporal dependencies. Our results emphasize the importance of incorporating climatic factors such as soil moisture, temperature, and precipitation in forecasting models. These findings contribute to the development of targeted pest control strategies tailored to specific agroclimatic zones and enhance forecasting and monitoring methods based on GIS and computational modeling